What I build. What it does and what it delivers.
This is what I build, measure and use every day myself. I do not write publicly about client work.
Finished and in use
Working, and in use.
- Own hardware
- No cloud
- LLMs
my own AI server, no cloud
For your organisation: AI on your own hardware: your data stays in, and you do not depend on a cloud provider's prices and terms.
Runs the LLMs for my other projects on my own hardware, more than 3 times as fast as the well-known open-source engine on the same cards, and trains speech models too. A job scheduler queues GPU work, and a card that is free picks it up by itself.
Tech stack: An AMD EPYC 7443 with 24 cores and 256 GB of memory, two RTX 3090s at 250 W and a Radeon RX 6600 XT for the Mac VMs; Proxmox runs the virtual machines.
more than 3x as fast as llama.cpp
More about HydraClose
Hydra
What it does
My own models on my own hardware, more than 3 times as fast
Hydra runs, trains and tests the models for my other projects locally, without the cloud. The large language model runs on my own inference setup, on an open-source engine with my own changes, including a GPU kernel of my own (3 to 7% faster). That is more than 3 times as fast as llama.cpp with its best settings, the well-known open-source engine that tools like Ollama and LM Studio build on. Proxmox splits the machine into virtual machines; a job scheduler queues GPU work, and a card that is free picks up the next job by itself.
- for one conversation with Qwen 3.8 Flash Next (125 billion parameters), on short text and long code; up to 209 when the answer copies from the prompt
- 136 tokens/s
- as fast as llama.cpp with its best settings, the engine under Ollama and LM Studio, on the same two cards
- over 3x
- in total with four conversations at once
- about 200 tokens/s
- when reading in a prompt (up to 2,600); at 30,000 tokens the first word comes after about 12 s
- 2,500 tokens/s
My own measurement, 6 November 2026, on my own server, normal settings.
One conversation, tokens per second: llama.cpp against my own inference setup
Short prompt
- llama.cpp
- llama.cpp, n-gram
- Open-source engine
- + own kernel
30,000-token code prompt
- llama.cpp
- llama.cpp, n-gram
- Open-source engine
- + own kernel
Answer that copies from the prompt
- llama.cpp
- llama.cpp, n-gram
- Open-source engine
- + own kernel
My own measurement, 6 November 2026, on my own server, normal settings; median of 512 generated tokens per prompt, run-to-run noise about 6%. llama.cpp with the best offload over both cards, and with n-gram speculation. Open-source engine: with my own changes; + own kernel: adding a GPU kernel of my own, 3 to 7% faster, with answers just as good. Against llama.cpp with its best settings that is about 3.4x on short text, 4.5x on long code and 2.2x when the answer copies from the prompt. Several conversations at once: total tokens per second
No batching
- 1 conversation
- 2 conversations
- 4 conversations
Batched, up to 4 at once
- 1 conversation
- 2 conversations
- 4 conversations
My own measurement, 6 November 2026, on my own server, normal settings. Batched with four conversations: the best of three runs (192, 197 and 206), usually about 200. Batching turns more users into more throughput: with four conversations about 50 tokens per second each. Tokens per second: roughly words per second, much faster than you read. What runs where
Proxmox: the virtual machines
- AMD EPYC 744324 cores, 256 GB of memoryjob scheduler, virtual machines and the experts that are needed less often
- RTX 3090limited to 250 Wfirst part of the layers, plus the most-used experts
- RTX 3090limited to 250 Wsecond part of the layers, plus the most-used experts
- Radeon RX 6600 XTfor the Mac VMsdisplay for the virtual Macs
Schematic. The language model's layers are split across the two 3090s; the most-used experts live on the GPUs, the rest in system RAM, and 99.1 to 99.8% of the work hits the GPUs. The machine
What is inside
- Motherboard
- ASRock Rack ROMED8-2T, two 10 Gb ports
- Processor
- AMD EPYC 7443, 24 cores, Noctua NH-U14S
- Memory
- 8 × 32 GB ECC DDR4-2666, all 8 channels
- GPUs
- 2 × RTX 3090, each limited to 250 W
- Mac VMs
- AMD Radeon RX 6600 XT
- Power
- be quiet! Dark Power Pro 13, 1600 W
- Virtualisation
- Proxmox VE
- Management
- fully remote, like a real server
- Network
- 10 Gb
A smaller model, twice
Two coding assistants at once
For coding work a 27B model (Qwen 3.8 27B) runs on each 3090 separately: two conversations at once, each with up to about 237,000 tokens of context. Measured on real coding work between 22 October and 6 November 2026.
- per card on longer answers (median); short answers about 110
- 70 tokens/s
- when the answer copies a lot from the prompt
- up to 309 tokens/s
- of real coding work in just over two weeks
- over 6,000 requests
Live log of the coding service, over 6,000 requests, 22 October to 6 November 2026.
Also for training
37 speech recognition adapters in one campaign
Hydra also trains models. For the Dutch speech recognition in Scribe, a campaign in August 2026 trained 37 adapters (fine-tuned add-ons to a model): for Whisper turbo and small, for Parakeet, and one probe with Whisper large-v3. What came of it, and why the app still runs stock Whisper, is under Scribe.
1 of 13- Open source
- AI agents
- npm and Homebrew
a fuel gauge for AI agents
For your organisation: a team that runs AI agents on subscriptions sees up front whether a job still fits, and does not stall halfway through.
Keeps an AI agent from stalling halfway through a job because the plan ran out. Headroom reads how much room is left in your Claude and Codex plans and tells the agent: start, wait or switch to another plan. Open source.
Tech stack: TypeScript; a command-line tool, a connector for AI assistants (MCP) and a Claude Code plugin. Installs with npm or Homebrew.
9,605 downloads in two months
More about HeadroomClose
Headroom
What it does
Go, wait or switch
Headroom reads the 5-hour and weekly limits of your Claude and Codex plans and gives an agent one answer: start, wait, or switch to a plan with room. Open source (MIT), installs with npm or Homebrew, as a command-line tool, as an MCP server (the connector AI assistants use to call a tool) and as a Claude Code plugin.
- since the first release on 5 September 2026, up to 5 November
- 9,605 npm downloads
- in the 30 days up to 5 November 2026
- 3,578 npm downloads
Download figures from npm for the package, 5 September to 5 November 2026, read on 6 November.
npm downloads per week
- 4 Sept
- 2,067
- 11 Sept
- 966
- 18 Sept
- 882
- 25 Sept
- 1,194
- 2 Oct
- 1,012
- 9 Oct
- 174
- 16 Oct
- 1,283
- 23 Oct
- 674
- 30 Oct
- 1,353
Per week, for the npm package headroomd, 4 September to 5 November 2026. npm counts every install, automated ones too. Homebrew installs go through npm and are included in this number. npm downloads, running total
- 4 Sept
- 2,067
- 11 Sept
- 3,033
- 18 Sept
- 3,915
- 25 Sept
- 5,109
- 2 Oct
- 6,121
- 9 Oct
- 6,295
- 16 Oct
- 7,578
- 23 Oct
- 8,252
- 30 Oct
- 9,605
The same, added up from 5 September: 9,605 on 5 November. 2 of 13PodcastD
- Local
- Python
- RSS
from conversation or document to podcast
For your organisation: a policy, report or weekly wrap-up as an episode colleagues listen to on the move, made on your own server.
Turns a conversation, document or link into an episode in my podcast app, voiced on my own computer. So I hear on the move what my AI agents have done, instead of reading it behind a screen.
Tech stack: Python; a language model (local, or Claude) writes the script, speech models on your own computer voice it, a private RSS feed delivers it.
In use every day
More about PodcastDClose
PodcastD
What it does
One command, listen on the move
PodcastD is a skill in Claude Code. With one command the current conversation, a document or a link becomes an episode with one or more voices in my podcast app. Mostly it is an update from AI agents that keep working on something for hours; also project updates, a weekly wrap-up or a paper. That way I keep up without sitting behind a screen. I use it every day and plan to open-source it at some point.
How it works
Voiced locally
A language model (local, or Claude) writes the script, three speech models on your own computer voice it, and the episode lands in a private feed your podcast app picks up. On the local path nothing leaves the machine.
3 of 13- Online
- No cookies
- Self-hosted
this site: fast, private, self-hosted
For your organisation: a site that loads fast on a phone and needs no cookie banner, on your own server and with no lock-in to a website platform.
Loads on a phone at less than a tenth of the weight of a typical home page, with no cookies and no third-party trackers. The site you are reading, in two languages. I count visitors with a tool of my own that sets no cookies.
Tech stack: Astro, with no JavaScript framework in the browser, on a small server of my own.
Lighthouse 100 for performance
More about martijnhermans.nlClose
martijnhermans.nl
What it is
A fast site with no cookies
Static pages in two languages, built with Astro. No JavaScript framework runs in the browser, only a few small scripts: for the theme, the form, the animation on the home page and my own visitor statistics without cookies. No third-party scripts, so no cookie banner.
- Lighthouse performance score, run of 4 November 2026
- 100
- for the home page on a phone; the typical mobile home page weighs 2,362 KB
- 223 KB
- cookies and third-party scripts
- 0
My own measurements: Lighthouse on 4 November 2026, page weight on 6 November 2026. The typical home page comes from the HTTP Archive Web Almanac 2025.
How it is built
- Text and datain plain files
- Astro buildsstatic pages
- A small server of my own
- Visitor statisticsown tool, no cookies
4 of 13Private chat
- Local
- MCP
- Open WebUI
chat with my own models, locally
For your organisation: a chat like ChatGPT for your team, on your own hardware, so questions and documents stay in.
A chat like ChatGPT, but on my own AI server. My own command-line tool around Open WebUI, the open-source chat interface I use, starts the big models on Hydra, shows which model is loaded and gives every model in the list a status dot: ready, loading or off. Everything runs locally, with context from my own knowledge base over MCP.
Tech stack: Open WebUI as the chat interface, a command-line tool of my own that starts the models on Hydra, and a connector (MCP) to my knowledge base.
In use
More about Private chatClose
Private chat
What it does
Pick a model, see whether it is ready
Every model in the list has a status dot: ready, loading or off. Pick a big model that is off, and the tool starts it on my own AI server and shows which model is loaded. The 27B coding model comes in two variants: a fast one, and a "deep" one with thinking on. Everything runs locally, with context from my own knowledge base over MCP. The chat itself is Open WebUI, an open-source chat interface.
Speeds and measurements are under Hydra.
How it works
- Pick a modeldot: ready, loading or off
- The tool starts iton my own AI server
- Chat in Open WebUIwith context from my knowledge base
5 of 13Research agent
- Local model
- MCP
- Python
links and questions to sources
For your organisation: research that lands in your own knowledge base with sources, instead of in scattered chats.
Puts what is worth keeping into my knowledge base with its source and date, and keeps the rest out. I give it a link or a question; a local model judges whether it fits and a script decides: keep, park or reject.
Tech stack: Only the Python standard library; a command-line tool and a connector for AI assistants (MCP).
In use
More about Research agentClose
Research agent
How it decides
Keep, park or reject
A working tool I use when I find something worth keeping. One command with a link fetches it. A local language model scores every link in a fixed schema: does it fit my projects (fit) and is it new (novelty)? A script then decides, not the model. Doubt goes to parked, never into the knowledge base.
Pick the model's scores, see what the script decides
Reject
To the reject list; this link is never fetched again.
Reject
To the reject list; this link is never fetched again.
Reject
To the reject list; this link is never fetched again.
Reject
To the reject list; this link is never fetched again.
Park
30 days in the parked folder, then it expires.
Park
30 days in the parked folder, then it expires.
Reject
To the reject list; this link is never fetched again.
Park
30 days in the parked folder, then it expires.
Keep
Into the knowledge base, with source URL, date, model used and status proposed.
This grid assumes a link with no flags. With the promo flag a link is always rejected; with any other flag it is never kept. If the model does not answer, or answers invalid JSON, the link is parked.
6 of 13- Small language model
- Fixed rules
- Coffee
an espresso coach, local on a phone
For your organisation: a product or help-desk adviser that answers from your own manuals, with fixed rules as a safety floor, and says honestly when it does not know.
An espresso coach that runs on the phone itself: a very small language model that is very good at one subject, because it gets the right context for every question, with fixed rules as a safety floor. Built out of a love of coffee (light-roast anaerobics, ideally).
Tech stack: A small language model on the phone itself; 20 fixed rules as a safety floor, with search over reference material and the model that writes the advice on top.
Local on a phone
More about VettaClose
Vetta
Why
A very small model, very good at one thing
How do you make a very small local model very smart on one subject? Vetta is my answer, for espresso, a subject I know well. It runs as an app on a phone, with the model in the phone's memory, and does well because it gets the right context for every question.
Status
In use, with no upkeep
In use since spring 2026, by me and a few friends, and it runs with no upkeep, locally. A broad release is not planned; I might open-source it later, like PodcastD.
More at vetta.coffee.
How it works
Three layers, with fixed rules at the bottom
Layer 1 is twenty fixed rules: predictable, under 50 ms, the safety floor. Layer 2 searches reference material (hybrid search plus reranking). On top, the language model writes the advice, and says when there is not enough information.
Try it: pick taste and time, see which fixed rule answers
Grind 2 clicks finer
Confidence: highrule R-01
Raise the temperature 1 °C
Confidence: mediumrule R-02
Grind 1 click finer and check the distribution in the basket
Confidence: mediumrule R-03
Keep your settings if you like it
Confidence: highrule R-07
If it still tastes flat: 0.5 g more coffee (R-08, confidence low).
Keep your settings if you like it
Confidence: highrule R-07
If it still tastes flat: 0.5 g more coffee (R-08, confidence low).
Keep your settings if you like it
Confidence: highrule R-07
If it still tastes flat: 0.5 g more coffee (R-08, confidence low).
Grind 1 click coarser and check the distribution
Confidence: mediumrule R-06
Lower the temperature 1 °C
Confidence: mediumrule R-05
Grind 2 clicks coarser
Confidence: highrule R-04
7 of 13- Photography
- Astro
- Admin page
photography as an online exhibition
For your organisation: a portfolio or catalogue that reads like an exhibition rather than a web shop, and that you update yourself.
My photographs in three rooms, each with its own text and a fixed order, with the stories behind the images. I edit texts and order on an admin page of its own.
Tech stack: Astro, static, with the story of each work in a plain text file.
Online
More about hermans.studioClose
hermans.studio
What it is
An exhibition, not a web shop
Photographs hang in three rooms, Order, Earth and Quiet, each with its own text and a fixed hang order. Next to them, two documentary series with the stories behind them. An admin page of its own edits the texts, the order and the cover of each room.
How it works
- Admin pagetext, order, cover
- Saved per room
- Static pagesper room and per work
8 of 13- Blog
- Markdown
- SVG charts
notes on AI, building and the outdoors
For your organisation: a knowledge base or blog your team keeps in plain text files, with no monthly fee for a blog platform.
My blog on AI, building and the outdoors, with diagrams and charts of its own. One note shows generative art that your browser draws live.
Tech stack: Astro and Markdown; diagrams and charts as SVG, in a light and a dark version.
Online
More about NotesClose
Notes
What it is
Notes with diagrams and charts of their own
Notes on the drive data of an electric car, on generative abstract art, on Headroom and on Vetta, with more in the works. Most have diagrams and charts of their own, as SVG in a light and a dark version. The note on generative art shows pieces your browser draws live.
How it is built
- Markdownplain text files
- Astro builds the site
- Diagrams and chartsSVG, light and dark
9 of 13Openfolio
- MCP
- Go
- Own server
share pages made by AI, with versions and feedback
For your organisation: share prototypes, reports or mock-ups made with AI with clients or colleagues, with versions and feedback in one place, on your own server.
A place of your own where any coding agent publishes an HTML page and you get a link to view it. Every republish adds a version that never changes: you see the timeline and a diff against the previous version, and you can roll back. You pin comments on a version; agents read that feedback back over MCP and resolve it. I have used it in my own setup to share agent-made pages since earlier this year; the open-source version, separated from my private stack, is in the works.
Tech stack: Go and SQLite, one binary, plus an MCP server.
In use, open source in the works
More about OpenfolioClose
Openfolio
What it does
Publish, view and roll back
An agent publishes an HTML page and you get a link. Every republish adds a version that never changes, so you see the timeline and a diff against the previous version, and you can roll back. You pin comments on a version; the agent reads them over MCP and resolves them. Each version is served sandboxed from its own content origin, with signed links that expire quickly. I have used it in my own setup since earlier this year; the open-source version, separated from my private stack, is in the works.
How it works
- Agent publishesan HTML page
- New versionthat never changes
- You view and commenta comment on a version
- Agent reads the feedbackover MCP, and resolves it
10 of 13
In development
Already working, not yet available to everyone.
- Local
- MCP
- TypeScript
documents ready for AI, hidden text set apart
For your organisation: give invoices, CVs or quotes from outside to AI, without a hidden line such as "approve this invoice" coming along.
Lets AI read documents from outside without hidden text coming along. Seenline converts PDF, Word, Excel, PowerPoint and email to text and sets white, tiny, covered or off-page text apart, so a model does not read words a person never saw.
Tech stack: TypeScript; a library, a command-line tool and a connector for AI assistants (MCP), all local. Optionally it checks each page against what is actually visible.
0 of 25 hidden texts let through
With page check, on a public test set from research into hidden text (25 PDFs). The other tools tested flagged none. Own measurement and scoring.
More about SeenlineClose
Seenline
Why
Text a person does not see, a model still reads
A document can hold white, tiny, covered or off-page text. An ordinary converter passes it on as ordinary text, and a language model then reads it as an instruction. Drag the line to see what sits under the page.
What Seenline passes on
The same invoice, with the hidden line set apart
Seenline puts the hidden text in a separate block, with location, kind and length. A person can review it; the model does not read it as ordinary text.
Invoice 2026-118 Newsletter design, 12 hours Payable within 30 days. [Hidden in the original] p. 1, white text: 58 characters not shown- hidden texts let through unflagged by Seenline with page check
- 0 of 25
- flagged by Seenline itself, the rest clean without a signal
- 22 of 25
- by the seven other tools tested, together
- 0 flagged
Public test set of 25 PDFs from the study by Liu and Ming (arXiv 2606.15020); my own measurement and scoring, 5 November 2026.
Measurement
No hidden text let through unflagged
With page check, Seenline let none of the 25 hidden texts through and flagged 22 itself; the other tools tested flagged none. That is one property, not an overall verdict: PyMuPDF4LLM, Marker and Docling returned more files fully correct (16, 14 and 12; Seenline 3, because the 22 flagged count separately), and several tools dropped visible text without notice. The numbers are shown below each row. My own measurement on the tools tested; I wrote the scoring myself.
The test set
A public test set from research into hidden text
Not a standardised benchmark: 25 PDFs that each, in a different way, show a person something other than what a tool reads from the file.
From the study What Users See Is Not What Models Read: Split-View PDFs in Document-to-LLM Supply Chains by Side Liu and Jiang Ming (2026). The files are on Zenodo, under CC-BY-4.0.
11 of 13- Mac
- Local
- Dutch and English
meeting notes on your own Mac
For your organisation: minutes of meetings with clients or staff, without recordings or transcripts leaving the Mac.
You take part in the meeting instead of taking notes. Scribe turns a Dutch or English meeting into a transcript with speaker names and a summary in the language of the meeting, entirely on your Mac. No bot joins your call, and an hour-long meeting is transcribed in about ten minutes. Recordings and voiceprints stay there, and you can delete the voiceprints.
Tech stack: A Mac app (Electron, with a local Rust engine); Windows next. Speech recognition, summaries and speaker names run on the Mac itself.
In development
More about Modinus ScribeClose
Modinus Scribe
What it does
Stop taking notes, stay in the meeting
Scribe turns a Dutch or English meeting into a transcript with speaker names and a summary, in the language of the meeting. It records the microphone and the computer's sound directly on the Mac; no bot joins your call and nothing connects to the meeting. Everything runs on the Mac itself: speech recognition, summary and speaker names. Recordings stay on your Mac, and so do the voiceprints behind the speaker names; you can delete them. If part of the app crashes during a recording, the audio up to the crash is kept. Recording people needs their consent, so ask before you record.
Own speech recognition
Measure first, then build it in
I am fine-tuning my own Dutch speech model, on my own AI server. It makes far fewer mistakes with numbers, but it is not yet better across the board. So the app uses the standard Whisper turbo until my own model wins on every measure.
- until an hour-long meeting is transcribed, with speakers separated
- 10 min
- until the summary is ready, with decisions, actions and open questions
- 20 min
- outside the Mac while recording, transcribing and summarising: after installation Scribe works fully offline
- 0 connections
- misheard numbers in meeting-room recordings, with my own Dutch model
- 3.4 → 0.5 per hour
Measured by Scribe development with the installed app on a Mac mini (M4, 24 GB), 7 November 2026: a public 60-minute committee debate of the Dutch parliament, two runs; network measured per process during the whole run (the one-time model download at installation does use the internet). Numbers: measured through the app on Dutch recordings from the Spoken Dutch Corpus (CGN).
How it works
- Recordingthe Mac's microphone and sound, no bot
- Speech recognitionWhisper turbo, local
- Speaker namesvoiceprints stay on your Mac
- Summaryin the language of the meeting, on the Mac itself
12 of 13- Local
- Markdown
- MCP
your own knowledge, where your AI starts
For your organisation: your team and your AI work from the same maintained knowledge, so nobody has to explain a project over and over.
Modinus grew out of Atlas, my own second brain: my projects, decisions and sources in plain Markdown files that my AI agents read before they start and update when they are done. I have worked with it and built on it for a while. Modinus turns that idea into a product: your projects, cases and people in one place on your own computer, so your AI starts from what you already know.
Tech stack: Atlas is Markdown in a plain folder, plus a search index agents can query (MCP). Modinus runs on your own computer.
In development
More about ModinusClose
Modinus
Where it comes from
Atlas, my own second brain
Projects, decisions and sources live in plain Markdown files in a plain folder. My agents read them before they start and update them when they are done. A search index makes the folder searchable over MCP. I have worked with it and built on it for a while.
What it becomes
The same idea as a product
Modinus puts your projects, cases and people in one place, on your own computer, so your AI starts from what you already know instead of from zero. Modinus is in development.
How Atlas works
- Markdown in a folder
- Search indexsearchable over MCP
- Agents readbefore they start
- Agents updatewhen they are done
13 of 13